Research On Regularization Method Of The High Dimensional Tensor Of Seismic Signals

With the development and progress of modern society,the exploration and exploitation technologies for oil,natural gas and other energy resources have also been greatly developed.However,due to the development of exploration technology and the harsh detection conditions in the field,the seismic data collected during actual exploration are usually accompanied by a large number of defects.Incomplete seismic data can seriously affect the follow-up work of oil and gas production,which leads to the problem for reconstruction of seismic data widely concerned by scholars at home and abroad.And from the inherent low-rank,sparse and other characteristics of seismic data,many good optimization algorithms are proposed.However,with the continuous advancement of oil and gas exploration,underground structures in existing oil storage areas are becoming more and more complex.In order to more accurately detect the underground structure,while increasing the requirements for exploration level,higher requirements have also been put forward for the reconstruction performance of seismic signal reconstruction algorithms.In this paper,two different seismic signal reconstruction algorithms are proposed from the perspective of low rank prior constrain and data drive for poststack seismic data to improve seismic data reconstruction quality.1.This paper proposes a regularization method based on low rank tensor decomposition.This method uses a new tensor decomposition model to combine the low rank constraint with the Hankel transform reasonably.The low-rank tensor decomposition model itself has strict requirements on the low rank of the data,and the Hankel transformation just rightly enhances the low rank of the data.On this basis,this paper designs different algorithms for the reconstruction of noisy and no-noise poststack seismic data.Among them,the no-noise version algorithm NHAM can recover 4-5 orders of magnitude higher than the similar algorithm when the data's low rank satisfies the algorithm's requirement.The proposed algorithm greatly optimizes the speed of the algorithm while guaranteeing the reconstruction accuracy.The noisy version algorithm DHAM introduces a large number of matrix transformation and expansion operations in the solution process,making the solution speed relatively slow,but it can guarantee the reconstruction.At the same time improve the signal to noise ratio of the data.Thus,the recovery strategy can be adjusted according to different requirements to achieve an ideal reconstruction effect.2.Due to the fact that the existing algorithm model with low rank constrained regularization can not completely fit the characteristics of seismic data itself,this paper proposes a new generation countermeasure network structure: 3D tensor completion network.Starting from the seismic data itself,the network fits the mapping relationship between the missing tensor and the complete tensor pair through iterative training,and achieves the purpose of obtaining the reconstruction result by inputting the missing data.In order to make full use of the high-dimensional spatial information of post-stack seismic data to reconstruct seismic data,a three-dimensional convolution kernel operation was introduced into the network structure to achieve the purpose of tensor completion.And through the Adam algorithm to optimize the network parameters,batch normalize and other methods to optimize the network parameter convergence speed.After completing large-scale sample data training,the complementary network,because the model abstract enough seismic data features,even if the data body itself does not meet the lowrank model of the low-rank constraints,but also through the observation data is very good Create a complete data body.